chelydra
/
metal-kompanion
4
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

ki: implement multi-text embeddings in OllamaProvider::embed by batching one request per text and aggregating results

This commit is contained in:
Χγφτ Kompanion 2025-10-19 08:33:27 +02:00
parent e1eb3afacc
commit 2ee757b64a
1 changed files with 34 additions and 34 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

64
src/KI/Provider/OllamaProvider.cpp
View File

@ -108,43 +108,43 @@ QFuture<KIReply*> OllamaProvider::chat(const KIThread& thread, const KIChatOptio
QFuture<KIEmbeddingResult> OllamaProvider::embed(const QStringList& texts, const KIEmbedOptions& opts) QFuture<KIEmbeddingResult> OllamaProvider::embed(const QStringList& texts, const KIEmbedOptions& opts)
{ {
QNetworkRequest req{QUrl(QStringLiteral("http://localhost:11434/api/embeddings"))}; // Execute one request per input text; aggregate outputs.
QFutureInterface<KIEmbeddingResult> fi;
fi.reportStarted();
if (texts.isEmpty()) { KIEmbeddingResult r; r.model = opts.model; fi.reportResult(r); fi.reportFinished(); return fi.future(); }
struct Accum { QVector<QVector<float>> vectors; int remaining = 0; QString model; };
auto acc = new Accum();
acc->vectors.resize(texts.size());
acc->remaining = texts.size();
const QUrl url(QStringLiteral("http://localhost:11434/api/embeddings"));
for (int i = 0; i < texts.size(); ++i) {
QNetworkRequest req{url};
req.setHeader(QNetworkRequest::ContentTypeHeader, QStringLiteral("application/json")); req.setHeader(QNetworkRequest::ContentTypeHeader, QStringLiteral("application/json"));
const QJsonObject body{ {QStringLiteral("model"), opts.model}, {QStringLiteral("prompt"), texts[i]} };
QJsonObject data; auto rep = m_manager->post(req, QJsonDocument(body).toJson());
data["model"] = opts.model; connect(rep, &QNetworkReply::finished, this, [rep, i, acc, fi]() mutable {
data["prompt"] = texts.join("\n"); // Join all texts into a single prompt if (rep->error() == QNetworkReply::NoError) {
const auto obj = QJsonDocument::fromJson(rep->readAll()).object();
auto netReply = m_manager->post(req, QJsonDocument(data).toJson()); if (acc->model.isEmpty()) acc->model = obj.value(QStringLiteral("model")).toString();
const auto arr = obj.value(QStringLiteral("embedding")).toArray();
QFutureInterface<KIEmbeddingResult> interface; QVector<float> vec; vec.reserve(arr.size());
interface.reportStarted(); for (const auto &v : arr) vec.push_back(static_cast<float>(v.toDouble()));
acc->vectors[i] = std::move(vec);
connect(netReply, &QNetworkReply::finished, this, [netReply, interface]() mutable {
if (netReply->error() != QNetworkReply::NoError) {
// TODO: Handle error
interface.reportFinished();
netReply->deleteLater();
return;
} }
rep->deleteLater();
const auto json = QJsonDocument::fromJson(netReply->readAll()); acc->remaining -= 1;
const auto embeddingArray = json["embedding"].toArray(); if (acc->remaining == 0) {
KIEmbeddingResult res; res.vectors = std::move(acc->vectors); res.model = acc->model;
KIEmbeddingResult result; fi.reportResult(res);
QVector<float> embedding; fi.reportFinished();
for (const QJsonValue &value : embeddingArray) { delete acc;
embedding.push_back(value.toDouble());
} }
result.vectors.push_back(embedding);
result.model = json["model"].toString();
interface.reportResult(result);
interface.reportFinished();
netReply->deleteLater();
}); });
}
return interface.future(); return fi.future();
} }
void OllamaProvider::cancel(quint64 requestId) void OllamaProvider::cancel(quint64 requestId)